Research map / Detection and segmentation / CNN features
Less is More: Focus Attention for Efficient DETR
Dehua Zheng, Wenhui Dong, Hailin Hu, Xinghao Chen, Yunhe Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Sparse DETR token pruning with Focus-DETR on DINO | DINO · COCO val2017, keep ratio 0.3 | AP: 48.2 -> 50.4 | Improved |
| use Focus-DETR instead of full DINO | DINO · COCO val2017 | AP and compute: AP 50.9 -> 50.4; GFLOPs 279 -> 154; FPS 14.2 -> 20.0 | Mixed |
| apply Sparse DETR to DETR variants with learnable queries | Deformable DETR, DN-Deformable-DETR, DINO · COCO val2017 | AP: -1.9, -1.2, -2.7 AP | Hurt |
| add supervision with improved label assignment to foreground token selector (ablation) | Focus-DETR · COCO val2017, ResNet-50 | AP: 47.8 -> 48.8 | Improved |
| add top-down score modulation (ablation) | Focus-DETR · COCO val2017, ResNet-50 | AP: 48.8 -> 49.2 | Improved |
| add cascade token selection (ablation) | Focus-DETR · COCO val2017, ResNet-50 | AP: 49.2 -> 49.7 | Improved |
| add dual attention encoder (ablation) | Focus-DETR · COCO val2017, ResNet-50 | AP: 49.7 -> 50.4 | Improved |
| use top-down instead of bottom-up score modulation (ablation) | Focus-DETR · COCO val2017, ResNet-50 | AP: 50.2 -> 50.4 | Improved |
| add dual attention to models using Sparse DETR scoring without Focus-DETR scoring (ablation) | Deformable DETR, DN-Deformable-DETR, DINO · COCO val2017 | AP: +0.3, 0.0, -0.1, +0.3 AP | No clear effect |
| increase keep ratio from 0.3 to 0.5 (ablation) | Focus-DETR · COCO val2017, ResNet-50 | AP: 50.4 -> 50.5 | No clear effect |
Closest papers by meaning
- Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity (ICLR 2022 · similarity 0.91)
- DETR Does Not Need Multi-Scale or Locality Design (ICCV 2023 · similarity 0.89)
- Salience DETR: Enhancing Detection Transformer with Hierarchical Salience Filtering Refinement (CVPR 2024 · similarity 0.88)
- Hybrid Proposal Refiner: Revisiting DETR Series from the Faster R-CNN Perspective (CVPR 2024 · similarity 0.87)
- Not All Tokens Matter All The Time: Dynamic Token Aggregation Towards Efficient Detection Transformers (ICML 2025 · similarity 0.87)
- SpargeAttention: Accurate and Training-free Sparse Attention Accelerating Any Model Inference (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
